The concept of classification through deep learning is to build a model that skillfully separates closely-related images dataset into different classes because of diminutive but continuous variations that took place in physical systems over time and effect substantially. This study has made ozone depletion identification through classification using Faster Region-Based Convolutional Neural Network (F-RCNN). The main advantage of F-RCNN is to accumulate the bounding boxes on images to differentiate the depleted and non-depleted regions. Furthermore, image classification’s primary goal is to accurately predict each minutely varied case’s targeted classes in the dataset based on ozone saturation. The permanent changes in climate are of serious concern. The leading causes beyond these destructive variations are ozone layer depletion, greenhouse gas release, deforestation, pollution, water resources contamination, and UV radiation. This research focuses on the prediction by identifying the ozone layer depletion because it causes many health issues, e.g., skin cancer, damage to marine life, crops damage, and impacts on living being’s immune systems. We have tried to classify the ozone images dataset into two major classes, depleted and non-depleted regions, to extract the required persuading features through F-RCNN. Furthermore, CNN has been used for feature extraction in the existing literature, and those extricated diverse RoIs are passed on to the CNN for grouping purposes. It is difficult to manage and differentiate those RoIs after grouping that negatively affects the gathered results. The classification outcomes through F-RCNN approach are proficient and demonstrate that general accuracy lies between 91% to 93% in identifying climate variation through ozone concentration classification, whether the region in the image under consideration is depleted or non-depleted. Our proposed model presented 93% accuracy, and it outperforms the prevailing techniques.
In 1913, the French physicists Charles Fabry and Henri Buisson discovered the ozone layer. The inorganic compound with the chemical formula O3 is ozone or trioxygen [
In 1930, the British scientist Sydney Chapman identified the photochemical processes that eventually led to the ozone layer. The ozone layer or ozone protection is a stratosphere zone of the earth that blocks the bulk of the ultraviolet radiation of the sun [
Ozone properties were explored intimately by experimentalist Gordon Miller Bourne Dobson, who established an easy photometer that enabled the relative intensity at two wavelengths to be measured. Dobson also invented a gas O2 that combines to form the O3. The two atoms of oxygen are splits due to the absorption of light and then combine to make the ozone with the third molecule of oxygen. This technique is comprehended as photolysis. These artificial compounds usually occur inside the climate in tiny groups. While associating with an uncontaminated environment, there is a harmony between quantities of gas being made and the ratio of gas being comparable. Thus, the full concentration of gas inside the layer remains relatively consistent [
Ozone is a molecule made up of three oxygen atoms (O3) that exists naturally in the stratosphere (topmost atmospheric layer) in small amounts, helping to protect the existence of living beings on earth from the ultraviolet (UV) radiation of the sun. In urban areas, metropolitan cities such as Lahore and Karachi in Pakistan, the thinning of the stratospheric ozone layer is the most significant environmental issue [
Climate variation is demonstrated as a transformation either within the climate’s mean state or its nonstop variability for an associated period, for decades or longer. The consequences of global climate change are increasing quickly on the economic development of a country like. The Islamic Republic of Pakistan that already have low financial gain and poor human development indicators [
The ozonosphere could be a coating in soil’s stratosphere that comprises comparatively extraordinary absorptions of gas O3. This layer absorbs 93%–99% of the sun’s high-frequency actinic radiation, which is probably damaging to life on Earth. Over ninety-one of the gas in soil’s stratosphere is a gift here. It is mainly situated within the lower portion of the layer from around ten klicks to fifty klicks on top of Earth. However, the thickness differs seasonally and geologically [
Creating a precise prediction model involves collecting input parameters that cause ozone concentration changes [
Various approaches have been implemented to forecast ozone depletion worldwide, such as the deterministic approach and statistical regression. Few approaches have been built based on the principle of linear regression. Multiple linear regression (MLR) is amongst the most prominent linear regression models for estimating ozone depletion. However, the aforementioned models showed a downside in understanding the physical systems’ nonlinearity and complexity [
To provide a stable prediction process, developments in the neural network model have also been made over the years. By only using meteorological parameters as input data, recurrent architecture increases the efficiency of neural predictors dramatically. Besides, recent research to build a hybrid model focused on a convolutional neural network and long short-term memory (CNN-LSTM) has given an adequate statistical stable hybrid model, and its prediction efficiency has proven to be favorable to a multilayer perceptron (MLP) and long short-term memory (LSTM) [
In artificial intelligence (AI) terms, Faster Region-based Convolutional Neural Network (F-RCNN) is an advanced form of Convolutional Neural Network (CNN). It is a deep learning algorithm used to detect the images and classify them according to their features [
The identified issue is a cross-disciplinary topic that falls under a single big umbrella that contains physics, chemistry, mathematics, and computer science. The nonlinearity, complexity, and heterogeneity of the physical systems are sizable barriers in developing trustworthy prediction models for ozone depletion identification based on machine learning or deep learning techniques. It is desirable to use artificial intelligence-based approaches to solve these cross-disciplinary constraints that are considerably powerful with minimal processing time and computing resources to detect ozone depletion areas precisely with improved performance metrics. In fact, in the existing literature, CNN has been used for feature extraction. In that procedure, extricated diverse RoIs passed on to the CNN for the grouping purpose. However, the issue with this procedure is that various images have various prompting areas inside the picture, and it is difficult to manage and differentiate these RoIs that negatively affect the gathered results.
Various approaches have been implemented to forecast ozone depletion worldwide, but this research will contribute by achieving the following objectives:
Due to the unavailability of well-organized data source, identification of few reliable input parameters that cause ozone depletion. The creation of a precise prediction model involves identified input parameters. To settle the identified problem by limiting the number of areas in R-CNN. To achieve better results than existing techniques with the help of the proposed approach by fine-tuning of F-RCNN.
This paper is organized as follows. The introduction section presents a brief overview of ozone, the ozone layer, causes of depletion, effects on the ecosystem, developments made in image processing, and deep learning for the precise classification techniques previously used to identify ozone depletion. Section I tackles the problem statement and objectives. The literature review is clarified in section II. Material and methods have been discussed in section III. Validation and results have been represented in section IV. Finally, discussion about the study and results have been discussed in section V.
Machine learning and deep learning frameworks have shown massive upgrades in structure and aim to overcome performance-related issues by creating advanced classification models. Physical systems are chaotic in nature, due to which a small change in its input parameters can cause massive transformation in output. Therefore, this section sheds light on the literature regarding different causes of ozone depletion and techniques for the identification of depleted regions in the stratosphere.
In 2019, scientists compared all the regions of China and noticed that North China Plane (NCP) was more affected by air pollution. According to this research, the author observed that there was less research in the past on tropospheric ozone identification and decided to observe nitrogen dioxide NO2, carbon monoxide (CO), sulfur dioxide (SO2), and formaldehyde (HCHO) to determine the density of the troposphere in a vertical column, i.e., dynamical process EI Nino-Southern Oscillation (ENSO), potential vorticity (PV), quasi-biennial oscillation (QBO) and East Asian Summer Monsoon Index (EASMI). The author’s results predicted that troposphere O3 contains changes from increasing to decreasing tendency throughout summer 2005–2016. The author observed all the gas concentrations through satellite and found that these gases’ emission causes a decrease in troposphere O3. The author built a model to analyze the chemical process and dynamic process variables and their effects on troposphere ozone. As a result, the chemical process variable has a 69.73% trend, and the dynamic process variable has a 60.64% observed trend on troposphere ozone in NCP [
In 2019, researchers presented a study on troposphere ozone estimation for urban areas [
In the same year, estimation was performed in Quito to evaluate the concentration of the ozone layer. The estimation was done by using LUR models with PLS regression. LUR models were used to estimate the concentration of the ozone layer in urban areas [
In 2018, researchers observed the ozone layer’s surface in China in the past five years (2013–2017) to show severe pollution and its locally flexible tendencies. They resolved the effects of climatological distinctions on ozone through the linear regression model. The primary purpose of that regression model showed increasing ozone tendencies in megacity clusters of China. According to the observations, southern China was considered an area where the ozone depletion rate was high. The estimated rate of
In the same year, the authors developed a model based on ozone concentration using machine learning. Many studies have already been done on climate variation identification, e.g., global warming. Simultaneously, the estimation of overall changes in the ozone layer through fully interactive atmospheric chemistry was expensive. In this research, the authors propose a new technique using a simple linear machine learning-based regression algorithm to calculate the ozone layer absorption. The regression technique helps to calculate the three-dimensional ozone absorption from daily to monthly time intervals. Furthermore, the authors have highlighted the key steps to improve the range of machine learning-based ozone parameters [
Researchers in this study presented the chronological subtleties of ground-level ozone and its impacts on Kazakhstan. Due to the increasing air pollution tendencies and their effects on human beings, the authors compared the two cities’ ozone levels. The selected two cities were Almaty and Astana, to figure out the effects of air pollution on the environment. The study was conducted for the concentration of ozone in Almaty on an annual and daily basis and predicted that increasing ozone concentration trends caused severe impacts on the respiratory system in Almaty city. The ozone level became high in a warmer environment, and the emission of gases increased automatically. According to that research, the rate of respiratory diseases in Almaty city was higher than in Astana [
Scientists in 2019 worked on increasing trends of stratosphere ozone concentration. It has been observed that the stratosphere ozone is decreasing continuously over the 1998–2016 period. In this research, a space-based analysis of ozone layer concentration was performed. While in space, ozone is decreasing in the hot and even in Northern Hemisphere above 90% approximately. This decrease in ozone layer concentration did not disclose the Montreal Protocol’s ineffectiveness and discussed its positive effects on ozone recovery [
A research group in 2019 established a thermodynamic model to measure the concentration of the ozone layer at a low temperature for the first time. The researchers developed a thermodynamic model that measures the ozone layer’s yield when the final gas temperature was equal to the initial temperature and observed that ozone yield increased when the oxygen gas ratio was augmented [
Researchers in 2019 predicted ozone layer depletion through remote sensing algorithms. They used two climatologies in this research: Aura Microwave Limb Sounder (MLS) and the other is the Aura Ozone Monitoring Instrument (OMI). These climatologists monitor the ozone concentration every month with longitude (15
Four models with six distinct reference variables were explored in this study: Boosted Decision Tree Regression, Neural Network with Gaussian Normalizer, Neural Network with Min-Max Normalizer, and Linear Regression. To choose the correct input parameters for the proposed method, the Pearson Correlation Coefficient was employed. These models were then evaluated in Malaysia at three separate sites. The validity of the proposed model tested by estimating ozone concentrations for various time horizons, 24 hours, and 12 hours, has been examined in two different situations. The outcomes of the models are then compared by adding different performances to test their effectiveness. Finally, concerning the best model, the inconsistencies are calculated using the 95PPU and d-factor bracket indices [
A comprehensive discussion of a few prior similar research related to the proposed model’s traits entailed the classification system development based on the ozone concentration.
Our research supports climate variation management based on identifying the ozone layer’s concentration to predict the depleted and non-depleted regions through Faster Region-based Convolutional Neural Network (F-RCNN). In the existing literature, authors have used Convolution Neural Network (CNN) for feature extraction, where computer vision has become a fascinating field. In that procedure, extricate diverse Regions of Interest (RoI) and pass them to CNN for the grouping purpose. However, the issue with this procedure is that various images have various prompting areas inside the picture that have been extricated in this technique, and it is difficult to manage and differentiate those RoIs that negatively affect the gathered measures.
To solve this issue, a Region-based Convolutional Neural Network (R-CNN) is used that settles the identified issue by limiting the number of areas. R-CNN utilizes a specific search technique to choose a fixed number of RoIs (2000) and considers these regions as the suggested regions. R-CNN’s issue was that it requires some more computing power for preparation and organization due to the bulky grouping of 2000 RoIs for each picture and then learning the patterns. Faster R-CNN is presented to regulate the planning and computational issues; rather than taking care of the considerable number of RoIs to CNN, an information picture is passed to the CNN for the component maps. Highlighted maps help recognize the area of interest; after distinguishing the proposed areas, these areas are wrapped into the square bounding boxes that limit the RoIs. That is why we have used Faster RCNN to predict the depleted and non-depleted regions with bounding boxes.
To completely associate the layer with the corresponding regions, RoI Pooling is utilized to fix their estimate and reshape them for taking care of these regions. At that point, the Softmax layer is utilized in the prospect of class name against each proposed area, and it likewise gives the balance estimations of every region. However, Faster R-CNN’s issue is that when the proposed areas are utilized, they significantly hinder the computing process calculation. The downside of both R-CNN and Faster R-CNN is that they utilize a particular pursuit approach, a tedious method that influences the demonstration of the results [
Finally, the areas are in order, and every area’s counterbalance estimation is projected [
We have applied F-RCNN in place of a backbone strategy. It consists of various estimated kernel filters. It has eleven channels in the central convolution layer and five channels in the consequent convolution layer, having 3
RPN gives different proposed regions when it is applied to the images. We can say these proposed regions as CNN feature map sizes. It is an exciting task to work on feature maps having different sizes. To solve this issue, RPN has evolved to map these features and reduce the feature map with the same sizes. However, we know that the max-pooling layer has the same size, but the RoI pooling layer split the feature maps into equivalent regions. After completing RoI pooling, there are three other fully connected layers with 4096 channels in the first two layers. The third layer is responsible for classifying six different techniques for every class of the produced dataset. Thus, we can say that it has six channels for each class. In the end, the SoftMax layer plays an essential role in receiving each class of the dataset. Fully connected always has the same formation in all the systems. Fixed linear components are applied to the hidden layers to add nonlinearity.
In our proposed work, the ozone dataset is taken as input. The dataset is in the form of images; both depleted and non-depleted areas are included. It is essential to normalize the dataset and then place it in a sequence. Furthermore, we have applied Faster RCNN, which consists of three steps, as shown in
Faster RCNN is presented by Ross Girshick, Shaoqing Ren, Kaiming He, and Jian Sun in 2015. It is a famous object detection architecture and advanced form of CNN. It is a deep learning algorithm used to classify and feature extraction of images after training and testing. It consists of three steps. The network architecture diagram for Faster RCNN is shown in
The first step is the processing through convolution layers. Five different layers work behind the scene. According to our scenario, these layers apply convolution, max pooling, and fully connected layers to classify ozone layer images. Filters are trained in this step to extract the appropriate features of ozone images to classify them as depleted and non-depleted areas. The pooling layer is applied to minimize the pixels with low values and consists of a reduced number of features, as shown in
The second step is Region Proposal Network (RPN), in which input images are passed to get the features’ map. The main advantage of RPN is to use weight sharing between the RPN backbone and the Faster RCNN indicator backbone. Bounding boxes are labeled for feature maps. The pooling layer takes that region of the ozone image whose feature mapping needs to be done. RPN is a deep learning algorithm and acts like a sliding window to extract ozone images’ features as depleted and non-depleted regions. When ozone images are passed to the fully connected layer, then their classification and regression branches are served through identified features. These features are then passed through the SoftMax layer to get it done the classification process. Here anchors are represented as boxes, and they can be 3
A fully connected neural network is used to predict the classes and bounding boxes. The ozone images dataset is then learned by the Faster RCNN independently to get the results and accuracy measures. It is used to classify the images into two categories; SoftMax and regression. SoftMax plays a vital role in classifying ozone images and placing the boxes to highlight their depleted and non-depleted regions.
In deep learning algorithms, SoftMax has a great place in the classification of images and appears in logistic regression. We apply SoftMax to obtain output through matrix multiplication. For example, we have a SoftMax, and we denote the vector of logistics as
Now applying the multivariate chain rule, the Jacobian of P(W) is
The Jacobian of g(W) is remaining, here g is a simple function, and calculating its Jacobian is easy. The Jacobian of g(W) has T rows and NT columns.
The weight matrix is linearized to vector length NT. So, we have to linearize in row-major order where all the rows follow each other in ascending order. Precisely W
So,
Now we have the following equation of index W to compute i, j.
The dot product between row DS and column D
The above equation shows that we can compute the SoftMax without actual Jacobian matrix multiplication and alternatively we can calculate the SoftMax through the following formula. The complete description of this formula is given in
Faster RCNN follows a supervised learning approach, and due to some certain constraints and limitations we collected 200 images for the ozone depletion identification. Then, a python script has been developed for the augmentation of dataset, which is the primary requirement of the proposed model for the better capacity building. Augmentation process increases the number of images from 200 to 500. Overall, implementation is done in python using YOLOv3 for Faster RCNN. The accuracy rate is increased after every iteration. For labeling all datasets, we used label-mg, which is a desktop application. We applied label-mg to distinguish between depleted and non-depleted regions. When we applied label-mg on images, it converts the JPEG and PNG images into the XML file. In the XML file, we have all the images, which are labeled by us plus our values against it. After labeling all the images, Faster RCNN implementation is done that it is used for the classification and detection of images at a high accuracy.
A dataset of ozone layer comprising depleted and non-depleted regions is generated on our own to feed the proposed algorithm. The dataset contains almost 2000 images generated by us by applying annotation on images to label images. Five hundred images are collected for each class, and we have a total of two classes as shown in
Classes types | Training images | Testing images |
---|---|---|
Depleted | 500 | 500 |
Non-Depleted | 500 | 500 |
First of all, we have uploaded 500 images for training purposes and obtained a 90% accuracy rate. To increase proposed model’s accuracy rate, dataset is tested in the second iteration and got accuracy 92% and 90% for depleted and non-depleted regions respectively. The popular loss function (Mean Square Error) is used for the calculation of system’s overall loss. In
The overall calculation of loss is shown in
Here t is the number of data points available from 1 to v, Sj is the actual label, Sjp is the predicted label. The total accuracy is 93% in our proposed system. Training and validation accuracy is displayed in
Yu Chen Wang used the SVM classifier with the self-generated dataset and achieved 91% accuracy. Also, Xinchen Wang used the Faster RCNN with a Pascal VOC dataset and achieved 90.6% accuracy. We evaluated and compared our results with different existing architectures. The comparison of our results with the architectures mentioned above is displayed in
Research | Methodology | Accuracy (%) | Dataset |
---|---|---|---|
Proposed | Faster RCNN | 93 | Self-generated |
Yu Chen Wang | SVM Classifier | 91 | Self-generated |
Xinchen Wang | Faster RCNN | 90 | Pascal VOC |
Faster R-CNN based detection and classification.
Our identified system Faster RCNN achieved a overall accuracy of 93% during testing by using self-generated dataset.
Finally, in the third iteration, the accuracy rate for both depleted and non-depleted regions has been increased. In this iteration, the depleted region’s accuracy is 93% and 91% for the non-depleted region.
The number of epochs for training and the validation are 200,000, at which reliable and significant results were obtained.
The classification goal in deep learning is to predict the identified classes accurately for every presented case in the specified domain [
In conclusion, the anticipated deep learning-based classification model F-RCNN is deployed to monitor ozone in the stratosphere by detecting depleted and non-depleted regions for future environment-related policies and the betterment of ecology. Dataset has been created that incorporates current stratosphere images of two types based on ozone depletion and non-depletion. The dataset contains 28 features and two classes. The python platform is utilized in this work as it can work with high certainty because of its capability of creating authentic outcomes that mirror a specific domain of intelligence. The deep learning algorithm-RCNN was applied to the self-generated dataset for classification in two groups precisely: that have different ozone concentration levels. It will help to characterize the normal ozone concentration to ensure proper monitoring of climate variation and slow down the adversative changes by making prolific policies to maintain appropriate ozone levels.
Due to the research scope, some common types of image datasets and parameters have been selected, but in the future more categories of innovative images and features in the stratosphere can be incorporated to better understand and cope with the issues. Furthermore, innovative algorithms can also be designed, or existing ones can be modified to detect a broader range of variations in the dataset in order to get better efficiency and accuracy.
Thanks to our colleagues & institutes, who supported us morally.